From a photo to a reviewed pantry
Open My pantry, choose a clear kitchen photo, optionally adjust the crop, and allow online analysis. The app suggests visible foods and labels uncertain identifications. Correct each food name and its preparation state, select the ones you recognize, then decide whether to add them to your pantry or keep them as food notes.
What a photograph can and cannot tell you
A clear photo can provide evidence of visible foods. It is not a complete ingredient list for a cooked dish. Oil, salt, sugar, blended ingredients and hidden fillings may not be visible. Similar cheeses, herbs, powders and products in opaque packages can be difficult or impossible to distinguish. The tool therefore does not estimate calories, grams, freshness or allergy safety. Read original labels and verify food safety yourself.
No automatic ingredient substitutions
The recognition vocabulary is not restricted to our recipe library. Foods outside the 49 ingredient labels can be retained as exact food notes. They do not increase recipe-match percentages until you explicitly choose a matching recipe ingredient. Canned chickpeas do not silently become dried chickpeas; generic pasta does not silently become whole-wheat pasta. Cooked meal components are not automatically treated as available pantry supplies.
Take a better photo
Use good light without glare, keep the camera steady, and separate ingredients where possible. Include legible packaging for products that look alike. Crop closer for a label or small ingredient, while remembering that anything outside the crop is excluded from analysis. The app accepts JPEG, PNG and WebP up to 20 MB and 48 megapixels, then prepares a metadata-free JPEG no larger than 1600 pixels on the longest edge, which is sharp enough for the AI and quick to send. Convert HEIC/HEIF files to JPEG before selection.
Your consent and data
Photos are processed online only after consent. The browser strips metadata by redrawing the photo, and the app does not persist the uploaded image or raw model output. The selected provider is shown before submission. Provider policies and security checks also apply; read the privacy details. Recognition requires an internet connection. Manual pantry selection, recipes and saved plans continue to work when the relevant site assets are available offline.
For the website owner
This release includes a Cloudflare Pages advanced-mode _worker.js endpoint. Enable a Workers AI binding named AI, set PHOTO_RECOGNITION_ENABLED=true, and configure a Turnstile site key and secret for your actual hostname. Redeploy after changing bindings. Alternatively, set PHOTO_PROVIDER=openai and an encrypted OPENAI_API_KEY. No API key belongs in browser code. Check /api/vision/status and complete a real-photo acceptance test before advertising the service as operational.
How reliability is evaluated
Software tests cover upload validation, response validation, explicit confirmation, mapping rules, unsafe inputs and error recovery. They do not establish a model's recognition accuracy. Evaluate a consented, labeled photo set separately: clear ingredients, similar-looking foods, cooked dishes, sealed packages, poor lighting and non-food images. Report incorrect additions and missed visible foods, not an invented confidence percentage. A model change should repeat that evaluation.